SciPost Submission Page
Optimal, fast, and robust inference of reionization-era cosmology with the 21cmPIE-INN
by Benedikt Schosser, Caroline Heneka, Tilman Plehn
Submission summary
| Authors (as registered SciPost users): | Tilman Plehn · Benedikt Schosser |
| Submission information | |
|---|---|
| Preprint Link: | https://arxiv.org/abs/2401.04174v3 (pdf) |
| Code repository: | https://github.com/cosmostatistics/21cm_pie |
| Date submitted: | 2025-03-21 09:32 |
| Submitted by: | Schosser, Benedikt |
| Submitted to: | SciPost Physics Core |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
| Approaches: | Computational, Phenomenological |
Abstract
Modern machine learning will allow for simulation-based inference from reionization-era 21cm observations at the Square Kilometre Array. Our framework combines a convolutional summary network and a conditional invertible network through a physics-inspired latent representation. It allows for an efficient and extremely fast determination of the posteriors of astrophysical and cosmological parameters, jointly with well-calibrated and on average unbiased summaries. The sensitivity to non-Gaussian information makes our method a promising alternative to the established power spectra.
List of changes
- Change "optimal" to "efficient" in the abstract, clarifying the point of 'optimal' summaries.
- In section 2.5 include the sentence:
If there is no violation then one can claim that the posterior is on average without bias and uses all the information from the summary. However, different biases in certain regions may cancel each other and the summary might not be optimal.
- In section 3.2 include "on average" in the following discussion:
Unlike standard MCMC analysis, where SBC is computationally not feasible, we can claim that our posteriors have the correct coverage and on average no bias.
- In the outlook, we clarify that optimality is reached by joint training, given the network architecture we choose:
The summary vector that links the two networks of the 21cmPIE-INN is initialized to the parameters of interest, but adapted by the joint training with the cINN to guarantee an optimal inference, for the given underlying network.
- In the outlook we clarify what we mean by optimal in the following sentence:
This speed, combined with the ability to capture non-Gaussian information and learn optimal summaries (meaning well-calibrated with SBC and on average unbiased), distinguishes this method from alternative inference approaches.